DiscoSign Wants Sign Language Translation to Actually Follow a Conversation

Key points

• DiscoSign is an Apple ML Research system that translates text into sign language glosses using discourse context.
• It handles spatial coreference, ensuring people and objects stay in consistent locations throughout a signed conversation.
• It recognises Question-Answer Clauses, a structural feature of sign language that sentence-level systems typically miss.
• The framework is built on a Large Language Model and is described as modular.

Sign language isn't a string of isolated sentences. Meaning builds across a conversation: a person introduced on the left stays on the left, and a question implies its answer in ways that reshape what follows. Apple ML Research's DiscoSign is built around that reality.

The system targets three discourse phenomena. One is spatial coreference resolution: in sign language, entities occupy consistent locations in the signing space throughout an exchange, and translating without tracking that consistency produces incoherent output. A second is the handling of Question-Answer Clauses (QACs), pseudocleft structures in which a question and its answer form a single linked unit. Sentence-level systems don't see the link; DiscoSign does.

The underlying architecture is a modular framework built on a Large Language Model, an AI system trained to process and generate text. Modular here means each discourse phenomenon gets its own dedicated component rather than being folded into a single undifferentiated model, which should make it easier to audit or extend.

How does this affect sign language users?

More coherent translations matter most when the content itself is complex: a legal explanation, a medical consultation, a classroom lecture. Sentence-by-sentence systems can stumble on exactly those contexts, where spatial references accumulate and QACs carry logical weight. DiscoSign's approach won't eliminate errors, but targeting the right failure modes is the correct place to start.

We first covered AI sign language translation when Google DeepMind's SL2T model brought ASL dictation to Pixel phones on 12 August 2026. That system moved signed input to text; DiscoSign works in the opposite direction, from text to sign language glosses (written representations of signs used as an intermediate step in translation pipelines). The two efforts together sketch the outlines of a two-way translation stack, though neither team is describing a joint effort.

Should developers of sign language technology take note?

Yes. The honest point here is that discourse awareness has been a known gap for years, and DiscoSign is notable less for the idea than for shipping a concrete framework that addresses it. Developers building similar pipelines should look at the modular architecture: isolating discourse components makes each one testable in a way that end-to-end models typically aren't.

The thing to watch isn't whether the approach is sound, it's whether the benchmark gains hold up when the system moves from controlled glosses to real-world, domain-varied text. That test hasn't been described publicly yet.